Jev 作者发布 12 页 PDF：把 Jev 做成 agent 的第二大脑

NO1ennn @N01ennn原推发布于 2026.09.27

在 X 上打开原推

holy sh*t. Jev Founder, Diogo Almeida just released a 12-page PDF how turned Jev into a second brain for AI agents a CLI for your Obsidian vault, a Claude Code plugin and a research paper. this is a 9-step blueprint on how to give your agents memory without paying an LLM to decide what's worth remembering: step 1 → split the work: Jev decides what to remember, code applies the rules, the LLM only writes the note or the answer step 2 → gate every capture with one call: save, todo, skip or review, plus is it private and how useful will it be in 30 days step 3 → keep only what clears the bar: save at value 1.5+ and confidence 0.6+, anything private never leaves your machine step 4 → let the LLM write one note, 5 to 12 lines, one claim. Jev never writes the body, or you end up with lossy compaction step 5 → link it: Jev compares the new note with its neighbors and answers duplicate, related, revises or contradicts. related at 0.7+ gets wikilinks both ways step 6 → read cheap, then judge: BM25 or embeddings pull the top 10 to 20, Jev keeps only what the question needs, 5 notes max go into context step 7 → clean up every night: duplicates get a merge proposal, old notes get superseded_by, contradictions get a conflict stub, and you keep or reject step 8 → don't build from zero: jev-second-brain runs on your vault, jevmem plugs into Claude Code, Jev-Mem is the paper if you're building agent memory yourself step 9 → check the receipts: jevmem hit 98.5% on save/skip at 0.30s while six frontier LLMs took 2.8 to 4.3s, and Jev-Mem scored 0.777 vs 0.700 on LoCoMo with memory built 6.6x faster your agent doesn't forget because the model is weak. it forgets because nothing ever decided what was worth keeping memory was never a storage problem. it's a decision problem, and decisions are exactly what Jev is built for stop giving your agent a bigger context window. give it a judge

这条是引用着另一条写的：x.com/n01ennn/status/2103542021071978601